Goswami Meetpuri

CricMind: High-Performance IPL Analytics & AI Simulator

IPL cricket data is scattered across dozens of sites — fragmented, ad-heavy, and requiring constant internet. Fans have no single, privacy-respecting tool to explore 19 seasons of ball-by-ball data interactively. No platform lets you ask natural questions like "Does winning the toss matter?" and get instant data-backed answers. We built CricMind — a fully offline, browser-based IPL analytics platform that processes 1,226 match files into interactive dashboards, a conversational AI assistant (CricAI), and a match simulator — zero APIs, zero tracking, zero server dependency. Process We sourced 1,226 Cricsheet ball-by-ball JSON files (IPL 2008–2026) and built a Node.js preprocessing pipeline that normalizes team names, computes player stats for 680+ cricketers, and generates phase analysis (Powerplay/Middle/Death), toss impact data, and H2H records — outputting one optimized JSON for the frontend. We chose React 19 + TypeScript + Vite for an offline-first SPA. Charts use Recharts, animations use Framer Motion. CricAI is a local NLP engine with 11 query handlers — not a cloud LLM — keeping everything in-browser. We iterated on design using glassmorphism, neon accents, and seamless video backgrounds. Results Processed 395,011 deliveries across 1,226 matches and 680+ players — all running offline in-browser with zero API calls. Key discoveries: toss winners only win ~50.6% of matches (myth busted), but chasing teams win 53.8% vs 44.3% batting first — a 9.5% delta. Death overs produce the highest run rate (9.54) despite being only 25% of an innings. Built a working CricAI chatbot handling H2H queries, player lookups, toss analysis, and season data — plus a Monte Carlo match simulator with confidence scores. Reflection I'd replace CricAI's rule-based NLP with a local edge LLM (like Gemma) for open-ended queries while staying offline. I'd add ball-by-ball match replay timelines using our 395K delivery dataset. The simulator needs a real Monte Carlo engine (10K+ scenarios) instead of weighted probability. I'd invest the first hour building a complete design token system before any pages — would've saved 40% time later. Finally, I'd ship as a PWA with service worker caching for permanent offline mobile access and add PNG/PDF export for sharing.